Re: [Linux-stm32] [PATCH 00/16] STM32 configure UART nodes for DMA

From: Valentin CARON
Date: Mon Nov 21 2022 - 03:49:12 EST


Hi Uwe,

We found the issue, thank you to have reported it.

stm32-usart driver was not tolerant to a probe defer from DMA when the earlycon is active.

You can find the patch here:
https://lore.kernel.org/lkml/20221118170602.1057863-1-valentin.caron@xxxxxxxxxxx/

Valentin

On 11/9/22 14:48, Amelie Delaunay wrote:
On 11/8/22 16:28, Marek Vasut wrote:
On 11/8/22 12:59, Uwe Kleine-König wrote:
On Fri, Feb 04, 2022 at 04:41:55PM +0100, Erwan LE RAY wrote:
On 2/4/22 2:22 PM, Alexandre TORGUE wrote:
Hi Ahmad

On 2/3/22 18:25, Ahmad Fatoum wrote:
Hello Erwan,

On 03.02.22 18:10, Erwan Le Ray wrote:
Add DMA configuration to UART nodes in stm32mp15x (SOC level) and
remove it at board level to keep current PIO behavior when needed.
For stm32-ed1 and stm32-dkx boards, UART4 (console) and UART7
(no HW flow control pin available) are kept in PIO mode, while USART3
is now configured in DMA mode.
UART4 (console UART) has to be kept in irq mode, as DMA support for
console has been removed from the driver by commit e359b4411c28
("serial: stm32: fix threaded interrupt handling").

Do I understand correctly that your first patch breaks consoles of
most/all boards, because they will briefly use DMA, which is refused
by the stm32-usart driver and then you add a patch for each board
to fix that breakage?

We have two solutions and both have pro/drawbacks. The first one (Erwan
ones, can break the boot if the patch is taken "alone". Your proposition
avoids this breakage but deletes a non define property (which is a bit
weird). However I prefer to keep a functional behavior, and keep Ahmad
proposition. Ahmad, just one question, dt-bindings check doesn't
complain about it ?

Cheers
Alex


Such intermittent breakage makes bisection a hassle. /delete-property/
is a no-op when the property doesn't exist, so you could move the first
patch to the very end to avoid intermittent breakage.

I also think that the driver's behavior is a bit harsh. I think it would
be better for the UART driver to print a warning and fall back to
PIO for console instead of outright refusing and rendering the system
silent. That's not mutually exclusive with your patch series here,
of course.

Cheers,
Ahmad


The driver implementation will consider the request to probe the UART
console in DMA mode as an error (-ENODEV), and will fallback this UART probe
in irq mode.

Whatever the patch ordering, the boot will never be broken. The board dt
patches aim to get a "proper" implementation, but from functional
perspective the driver will manage a request to probe an UART console in DMA
mode as an error and fall it back in irq mode.

I didn't debug this further yet, but my machine (with an out-of-tree
dts) fails to boot 6.1-rc4 without removing the dma properties from the
console UART. This is a bug isn't it? The same dts created a working
setup with stm32mp157.dtsi from 5.15 + kernel 5.15.

Hi Uwe,

Could you confirm earlycon is enabled on your setup?

Without earlycon, boot is ok, even with dma properties, at least on stm32mp157c-dk2.


I can debug this further, but maybe you know off-hand what the problem
is?

+CC Amelie, as this might be related to the DMA series that landed recently:

$ git log --oneline v5.18..v6.0 -- drivers/dma/stm32*

Hi Marek,

We haven't yet investigated the issue, and if latest DMA updates could explain why earlycon breaks the boot.


+TO Valentin, as he's now in charge of UART driver.
Valentin and I will investigate this issue.

Regards,
Amelie